What can go wrong

in FEA?
As finite-elem ent analysis spreads

to designers who

naay

lack formal training in numerical

proc edures, practltloners must ask whether the most

approp nate techniques are being

used-and whether they

are producing accurate results. By Klaus-Jiirgen Bathe

INITF,-EI,EMENT METHODS ATC lN

abundanr usc in today\ engineering prrcti,'e rhrough vari-

E5 with h very small = E

ous general-purpose commercial

computer programs and many special-pr-rrpose programs written for

specific applications. These techniques Jrc. to

ln

increasing extent.

being used to help identiS' good new

designs and improve designs with

respect to performance and cost.

Considering the important role

that finite-element methods now
play in various areas of engineering.
practitioners need to ask themselves

whether their procedures are the

most appropriate techniques available and whether the methods will

A schematic plot of strain energies Ep of finite-element solutions is given when the element size h is decreased. The plot shows convergence as the mesh is refined.

lead to accr-lrate results. These questions are particularly

important because more and more design engineers who
have not necessarily been trained in numerical procedures
are applying finite-element techniqr-res in their work.
As the use of these methods expands to a larger and
more diverse group, users must address the irnportant
question of what can go wrong in finite-element analysis.

This article is not intended to resolve the question in the

broadest sense; rather, we shal1 focus on some aspects
of the reliability of finite-element methods and their
accurate use. To illustrate. we will consider linear elastic
solutions and assume that the aigebraic finite-element

Klaus-liirgen Bathe is proJessor oJ mechanical engineeting

at the Massachusetts Institute oJ Technology in
Cambridge. He is also the Jouniler and a ilfueetor
of ADINA R&D Inc, in Watertowq, Mctss.

MRrnemRrlcAl MoDELS

equations are solved exactly. For a more cornplex analysis,

the sarne considerations ho1d, but additional requirements
need to be addressed as we1l.

First of all, engineers should recall that the finite-elerrent method is used to solve a mathematical model,
MECHANTcAL ENctNEEI{tNG Nl,tv

tcss

63

which is the result of an idealization of the actual physical problem considered. The mathematical model is
based on assumptions made regarding the geometry,
material conditions, loading, and displacement boundary
conditions. The governing equations of the mathematica1 model are in general partial differential equations subjected to boundary conditions. These equations cannot
be solved in closed analytical form. Therefore, engineers
resort to the finite-element method to obtain a numerica1 solution.
Consider, for example, an analysis of a valve housing of
axisymmetric geometry and axisymmetric loading. In
such a case, it is reasonable to assume axisymmetric conditions for analysis. The complete marhematical model
and thus the analysis problem is obtained by specifying
the geometry and dimensions, support conditions,
material constants, and
loading.

the mathematical model as /r approaches 0 for any (physically realistic) material data, displacement boundary conditions, and loading applied; and for a reasonable finiteeiement mesh, a reasonable finite-element solution will be
obtained. Furthermore, the quality of the finite-element
solution does not change drastically when the material
data (or thickness of a shell) are changed.
These conditions are of crucial importance. If the first
condition (convergence) is violated, then with rnesh refinement a solution is

approached that is not

the exact solution of thc

still in use could yield

grossly erroneous solutions.

mathematical model.

unique, and an approximation of this exact solution can

be obtained with very high accuracy using finite-element methods.

To quantify these observations, the notion of convergence must be introduced. Let E denote the (unknown)
exact strain energy of the mathematical model. Also, let
-87, denote the strain energy corresponding to the finiteelement solution with a mesh of element size ft. Then
convergence means:

lt-

Reliability in finite-element methods means that in the

solution of a well-posed mathematical model, the finiteelement procedures will have two attributes: The finiteelement solutions will converge to the exact solution of

Finite* element formulatio ns

While engineers cannot, in general, obtain

analytically the exact
solution for the posed
mathematical model in closed form, the exact solution
of the mathematical model does exist, the solution is

1T

THE QUESTtoN oF ReLtRerLlry

Ehl

The schematic piot on page 63 shows how convergence

is reached. As the mesh is refined (that is, the element size
h is decreased), the strain energy E1 approaches a value -E.
The rate at which the error between E and 81, is reduced
depends on the problem solved as well as on the element
and meshes used. Cleariy, higher*order elemenrs will reduce the error at a faster rate with mesh refinement than
lower-order elements.

l6-element
model
3x3Gauss
2x2causs
Mode
intesration
integration
1
LL2.4
110.5
2
634.5
617.8
3
906.9
905.5
4
1,548
958.4a
5
2,654
1,528
6
2'691
2'602
aspurious, i.e., phantom mode.

Such erroneous solutions

could lead to wrong design decisions and disastrous consequences. Of

course, finite-element methods that violate the first condition should not be used.
To consider the second conditi.on-obtaining a reasonable solution using a reasonable mesh-assume, for example, that the valve housing is made of steel (Young s
modulus is 200,000 megapascals; Poisson's ratio, V, is 0.30).
The analysis using a reasonable mesh has given accept-

error lE- Et l is acceptably smali).

Suppose that we now change the material to a piastic, for
which Poisson's ratio is 0.49-close to the incompressible condition of 0.50. This change in material condition
might result in a relatively small change in the exact solution, and it should then result only in the corresponding
able results (that is, the

small change

in the finite-element solution. lJnfortu-

nately, finite-element formulations are still used that violate

the second condition and yield a soiution grossly in error

when Poisson's ratio is changed to 0.49.
This solution phenomenon is observed with displacement-based finite elements. The large errors are present
because the elements are much too stiff when v is close
to 0.50, and cannot be used when V:0.50. The underlying mechanical reason is that,

considering the stresses, p:Ke,

16x64

-,
elementmodel

3x3Gauss
integration
110.6
60;:;
go',.2
L,44L
2,345
2'664

wherep is pressure, Kis bulk modr.rlur, ,od ej is volumetric strain. As

V approaches 0.50, K becomes
and is infinite when
"";;rl;r;"
: _^ '

v:0.50. Also, as v approaches 0.50,

in the exact solution, eu becomes
very smal1 and is zero when

0.50. Therefore, in an almost in_

cornp'essible analysis. the pressure

is

given by a very large number (the

bulk modulu$ multiplied by a very

,' ,.r'
,'
- rr
matrixwasused.Theresultsusingafinemesh(with64elementsreplacing"""rrri""-""d""1".""i"i small number (the volumetrlc
the l6-element mesh) are included for comparison purposes.
strain), and must be accurately calForthesixsmallestfrequencies(inhertzlofthebracketpresentedonpage65,theconsistentmass

64 uav 1998 MECHANICAL

ENGINTERTNG

nesh. As predicted bY theorY, using

full i.ntegration, the frequencies of

the coarse 16-element mesh are

larger than the accllrate solution
frequencies. When reduced integration is used, some frequencies
are better approximated than when
using full integration, but among
those few listed i.s a Phantom frequency. Phantom frequencies do
not physically exist but are only in50 mm

trocluced by the reduced-integration

scheme. If a dynamic step-by-step

N/mm2

solution is performed, such Phan-

E-55

tom frequencies are not noticed and

absorb energy, introducing large
errors into the solution. Error measures would detect that the errors
are large, but they are frequently not
available in dynamic analysis. Thus,
reduced integration is unreliable and

v=0-3
o:1.3.<10

'mm

N.sec2
e-

should be avoided.

Instead, only reliable mixed finite-eiement formulations, which

do not require reduced integration,
should be used. Such formr-rlations
satisfy the first and second conditions, are currently available, and
can be employed confidentlY for
the solution of mathematical models
with any material properties, loadi.ng, and boundary conditions. A
displacement/pressure-based formulation i.s particularly attractive
when incompressible materials are
The geometry and material data of a bracket used in an analysis are shown on top; the 16-element
mesh of nine-node elements used in the analysis is given on the bottom.

culated to balance the applied suface tractions. While the

bulk modulus becomes very large, the pressure is always
a finite number and usually does not change much as V
approaches 0.50.
As a remedy in displacement-based finite-element
methods, "reduced integration" is employed- This means

that in the numerical integration of the element stitTness

matrices, the exact matrices are not evaluated. The
method is simple to program and requires less computation tirne to establish the matrices, and with experience
acceptable results are frequently obtained. However, the
technique can also lead to very large errors.
Consider the frequency analysis of the bracket detailed
in the table on page 64 and the figure on page 65. In this
case, using nine-node elements, 3 X 3 Gauss integrrtion

corresponds to fu1l numerical integration tnd 2x2

Gauss integration is "reduced integration." Since no
analytical closed-form solution exists for the frcquencies
of the bracket, a very fine mesh (the 16 X 64 element
model) was used to obtain an accluate solution' Of interest are the solutions obtained

with the coarse 16-element

considered and was used to analYze

a plastic bracket with Poisson's ratio
equal to 0.499 (see Analysis Clinic,

March) . The formulation is also used effectively in

nonlinear analysis, where almost incompressible conditions are widely encountered. Inelastic conditions of
plasticity and creep, and rubberlike conditions. give rise

to (almost) incompressible behavior.

Similar difTiculties, as described for almost-inconlpressible analysis, are also encountered in the analysis of plate
and shell structures. Here, too, some finite-element technology still uses reduced integration, and the results can

in error. Again, reliable formulations,

which are now available, should be used instead of such
be very much

reduced-integration schemes.
To sum up, finite-element methods can now be employed with great confidence, but only the methods considered rehable should be used' Earlier technology based on
reduced integration should not be used, or should at any
rate be employed with great care. By proceeding in this
way, practitioners can have confidence that a finite-element
analysis will be effective and will not go wrong' r
Tht matcrial prescnted in thk artidt k tre ated at -futthct kngtlt in Finite Elenent
Procedures lty Klaus-lilrgen Bathe (Prentice Hall' 1996)
MLCTHANTcAL ENGINEERINcI Nr,q.v

9'lg

65